Observers say artificial intelligence often changes “how work is done” or “how well work is done” (quality improvements) rather than just “how fast it is done,” leading to outcomes that are difficult to capture in traditional productivity statistics.
That is particularly true for intangible products such as software, e-books, downloadable music, mobile applications, healthcare consultations, financial advice, legal services, streaming subscriptions, web hosting or any other product that is experiential (haircuts or live concerts).
Product quality changes, called “hedonic,” are particularly hard to quantify in these cases. Among the classic examples are personal computers that, over time, incorporate faster processors, more memory, better user interfaces, displays or audio, but without a price increase.
Smartphones might add premium materials, for example.
The point is that much of AI’s value is qualitative: improved decision-making, better user experiences, or reduced risk in complex processes (like drug discovery) that will not always show up as an immediate increase in volume-based output.
And all that is hard to measure.
Proxy Metric
What it Measures
Limitation
Task Completion Time
How much faster a specific, defined task is finished with AI.
Ignores quality variance and "rework" time (verification).
User/Adoption Rates
The percentage of the workforce actively using AI tools.
Does not measure value or net efficiency gains.
Resource Optimization
Reduction in compute or operational costs for a given output.
Can hide negative impacts on employee skill formation.
User Satisfaction
Improved quality of output or speed as perceived by the customer.
Subjective and may not correlate to bottom-line profitability.
Error/Defect Rates
Frequency of mistakes or need for human intervention in AI tasks.
Often hard to track consistently across different workflows.
That is not unusual for general-purpose technologies such as electricity or the internet. But financial analysts want quantitative metrics, so industries will develop them.
Industry
Metric Category
Specific Proxy Metric
Manufacturing
Operational Efficiency
Reduction in equipment downtime (via predictive maintenance).
Healthcare
Clinical Efficiency
Time reduction for diagnostic tasks or patient documentation.
Retail
Revenue & Customer
Increase in conversion rates or uplift in average order value.
Finance
Risk & Compliance
Reduction in fraud false-positive rates or manual audit hours.
Cross-Industry
Strategic Value
Revenue generated from AI-enabled new product lines.
Cross-Industry
Human Capital
Shift in employee time from routine tasks to high-margin work.
All of these metrics can be imprecise. It can be hard to isolate AI impact from all other organizational processes, for example.
Does use of artificial intelligence necessarily pose the risk of diminishing critical thinking or thinking skills? The answer might well depend on how AI is used.
But that is true of many human endeavors. People rarely stop engaging in activities they find intrinsically rewarding simply because technology makes the outcome easier to obtain.
As much as I enjoy this clip of waveriding at a favorite spot, I'd much rather be doing it.
Humans experience deep satisfaction when engaged in challenging activities that match their abilities. The reward is not merely the finished product but the experience of making it.
That is why people:
climb mountains despite vehicle access
bake bread despite supermarkets
garden despite grocery stores
play golf despite television coverage.
The activity itself provides enjoyment. If people only cared about outcomes, very few would actually play sports.
And AI may affect creative work in much the same way.
Calculators changed mathematics, for example. They reduced arithmetic effort but did not eliminate the need to formulate problems or interpret results.
There is a reasonable argument to be made that outsourcing “writing” to a language model poses a risk.
For many writers, that is a relatively negligible risk, since many writers compose because they enjoy the process of writing, and it makes no sense to outsource the “fun” of writing at all.
For many writers, writing is enjoyable because it combines:
discovery
exploration
self-expression
problem solving
craftsmanship.
The point is that technological advances rarely eliminate hobbies.
Photography did not eliminate painting
Recorded music did not eliminate amateur musicians
Power tools did not eliminate woodworking
GPS did not eliminate hiking
Word processors did not eliminate writing.
For many forms of knowledge work, AI is not replacing thinking so much as changing where the thinking occurs.
Instead of spending most of one's effort locating information, more of the cognitive effort shifts to:
Framing the problem before asking AI
Evaluating and challenging the response afterward
Synthesizing the results into an original conclusion.
So, in many research and knowledge-work settings, AI functions less as a replacement for thinking than as an accelerator for information retrieval and synthesis.
The important intellectual work frequently occurs before the AI interaction (defining the problem, framing the question) and after it (evaluating, integrating, and applying the results).
That is similar to how experienced researchers have long used search engines and databases, for example, and suggests that “how” AI is used matters.
And experts might benefit from using AI more than novices, as they are able to formulate better questions, based on:
relevant mental models
domain knowledge
intuition about good questions
ability to detect errors.
Stage
Traditional research
AI-assisted research
Where critical thinking occurs
Define question
Formulate hypothesis
Formulate prompt/problem
Very high
Gather information
Library, databases, Google
AI search or LLM
Moderate
Evaluate evidence
Read sources
Verify AI claims and sources
Very high
Compare viewpoints
Read competing authors
Ask AI to generate opposing views
Very high
Draw conclusions
Human synthesis
Human synthesis
Very high
Communicate findings
Human writing
Human writing (possibly AI-assisted editing)
High
Perhaps there is an analogy to the use of “search.” When Google became dominant around 2000, educators raised similar concerns about search making us dumber.
Instead, search shifted the balance of cognitive work away from memorization and toward higher-level reasoning.
AI might be similar, in many instances.
Thought traditional search asks users to perform much of the information synthesis themselves (find sources and evaluate them),
AI saves time by summarizing results.
But it does not eliminate the need to ask:
Is this correct?
Is something missing?
What assumptions underlie this answer?
What evidence contradicts it?
Are these the strongest sources?
Skeptics will note that many users will not take the time to do so. But that arguably was the case beforehand.
Rather than replacing reasoning, AI often expands the range of questions that can be explored within a fixed amount of time.
Also, there are some techniques that encourage broader and deeper exploration.
Recent research describes phenomena such as "cognitive offloading," "epistemic atrophy," and an "illusion of understanding," where users mistake fluent explanations for genuine comprehension. These effects appear strongest when AI substitutes for independent evaluation rather than supporting it. (Business Insider)
For experienced researchers, AI is often best understood as an unusually capable research assistant rather than an autonomous thinker.
But the researcher still bears responsibility for asking the right questions, verifying sources, weighing evidence, and integrating insights into an original conclusion.
In that sense, AI resembles an evolution of search rather than a replacement for thought.
The cognitive work shifts away from locating information and toward framing, evaluating, and synthesizing it.
AI substantially reduces the effort required for search, retrieval, summarization, and drafting, but arguably does not eliminate the need for problem formulation, judgment, skepticism, synthesis, or decision-making.
Whether critical thinking declines depends less on the technology than on whether users treat AI as an answer machine or as a research collaborator whose outputs require evaluation.
For writers who enjoy the process of writing, AI is not a replacement, anymore than watching surfing is a replacement for surfing.
A Gallup survey suggests 71 percent of Americans do not want a data center built where they live.
That “not in my backyard” (NIMBY) pattern occurs all the time, concerning landfills, homeless shelters, prisons, mental health facilities, wind turbines, airports, and cell phone towers, for example.
The dynamic reoccurs because of a “mismatch” between benefits and costs. When a policy’s benefits are broad and diffuse but its costs are concentrated, the people who bear the costs have the strongest incentive to organize, while the beneficiaries are often too scattered to mobilize.
That asymmetry helps explain why many useful policies and facilities trigger intense local resistance even when they are socially valuable overall.
In part, that is because the small number of recipients of the concentrated costs can mobilize easily. The perhaps millions of citizens who benefit are very hard to organize.
Politically, this is known as a “collective action” problem. The dispersed beneficiaries each have only a small personal stake, while the concentrated losers have a large one, so they show up at hearings, file lawsuits, and pressure officials far more effectively than the hard-to-organize majority.
So high-performance data centers are not “just” about technology: they also are about politics.
Data-center builders will need to engage in politics to get their projects built.
Depreciation schedules do not often assume strategic importance, but for neocloud suppliers of artificial intelligence “compute as a service,” depreciation of graphics processing units does seem to matter.
Assume a $12 billion investment in GPUs. If one assumes a three-year depreciation cycle, that produces a $4 billion per year hit to earnings.
On the other hand, if one assumes a longer six-year cycle, annual depreciation is just $2 billion per year.
The danger is the balance sheet hit if real-world useful life turns out to be less than six years.
Microsoft, Google, and Amazon arguably can absorb a bad depreciation call because they have robust other sources of revenue.
Neocloud providers must rely almost exclusively on the revenue from their GPU rental businesses.
So depreciation policy is an existential, not cosmetic issue. A useful-life error doesn't dent one segment's margin, it distorts the entire income statement, since:
Debt is often GPU-collateralized. Many neocloud financings are underwritten against assumed residual values. Most GPU financing deals assume a uniform, one-size-fits-all depreciation curve. If the real curve is steeper, the collateral coverage on that debt erodes faster than the loan amortizes.
Contract duration and useful life need to line up. If repayment schedules are structured around a six-year life but revenues fall after three years, debt servicing can become strained.
So schedule length maps to equity valuation. Longer schedules:
Lower annual depreciation, leading to higher reported net income and earnings per share.
Improve profitability optics today but risks painful impairments: if hardware is retired or written down early, the deferred expense hits all at once (an earnings "cliff") rather than being smoothed.
Skeptics argue that extended depreciation distorts actual operating metrics.
Shorter schedules also affect valuations:
Cause lower near-term margins and earnings per share, which can produce a valuation discount on trailing/forward P/E versus a peer using a longer schedule, even if the underlying cash economics are identical.
Lower restatement/impairment risk, and probably a lower cost of capital over time if investors reward accounting conservatism with a quality premium once the market re-prices this issue.
How much does it actually matter?
Some argue that since depreciation is a non-cash item, "the market sees through it."
Free cash flow is identical whether the schedule is three years or six, since the cash left the building at purchase.
A DCF-based or FCF-multiple-based valuation should be unaffected by the choice.
Others argue that generally accepted accounting practices still move equity prices. In this view, GAAP figures can be misleading because they're susceptible to noncash charges like depreciation, and metrics like EPS don't fully reflect a company's profitability.
So multiple-driven valuation can be distorted by the schedule choice, and equity-linked debt covenants (leverage ratios, EBITDA-based tests) can be gamed as well.
As always, the assumptions matter. Some argue GPU useful life really is in the three-year range.
In other words, even when a GPU cannot be used for training, it still has value for inference, other batch work or non-AI operations, generating revenue all the while.
Physical failure and retirement data for hyperscaler fleets do suggest older GPU generations last seven to nine years in production before physical retirement.
Company
Disclosed useful life
Posture
CoreWeave
~6 years
Most aggressive among pure-play neoclouds
Nebius
3-10 years (blended)
More conservative, cohort-dependent
Lambda Labs
~5 years
Middle ground
Microsoft / Oracle
Extended from 3-4 to ~5-6 years
Matches neocloud aggressiveness
Amazon (AWS)
~4 years
More conservative among hyperscalers
Meta
Up to 11-12 years in places
Outlier on the long end
Either way, the industry seems to be settling on a six-year depreciation cycle for GPU hardware.
